WIP:crepe

server/voice_changer/DiffusionSVC/pitchExtractor/CrepePitchExtractor.py

@@ -11,30 +11,26 @@ class CrepePitchExtractor(PitchExtractor):
     def __init__(self):
         super().__init__()
         self.pitchExtractorType: PitchExtractorType = "crepe"
+        self.f0_min = 50
+        self.f0_max = 1100
+        self.sapmle_rate = 16000
+        self.uv_interp = True
         if torch.cuda.is_available():
             self.device = torch.device("cuda:" + str(torch.cuda.current_device()))
         else:
             self.device = torch.device("cpu")
 
-    def extract(self, audio, pitchf, f0_up_key, sr, window, silence_front=0):
-        n_frames = int(len(audio) // window) + 1
-        start_frame = int(silence_front * sr / window)
-        real_silence_front = start_frame * window / sr
-
-        silence_front_offset = int(np.round(real_silence_front * sr))
-        audio = audio[silence_front_offset:]
-
-        f0_min = 50
-        f0_max = 1100
-        f0_mel_min = 1127 * np.log(1 + f0_min / 700)
-        f0_mel_max = 1127 * np.log(1 + f0_max / 700)
+    def extract(self, audio: torch.Tensor, pitch, f0_up_key, window, silence_front=0):
+        start_frame = int(silence_front * self.sapmle_rate / window)
+        real_silence_front = start_frame * window / self.sapmle_rate
+        audio = audio[int(np.round(real_silence_front * self.sapmle_rate)):]
         
         f0, pd = torchcrepe.predict(
             audio.unsqueeze(0),
-            sr,
+            self.sapmle_rate,
             hop_length=window,
-            fmin=f0_min,
-            fmax=f0_max,
+            fmin=self.f0_min,
+            fmax=self.f0_max,
             # model="tiny",
             model="full",
             batch_size=256,
@@ -46,14 +42,15 @@ class CrepePitchExtractor(PitchExtractor):
         pd = torchcrepe.filter.median(pd, 3)
         f0[pd < 0.1] = 0
         f0 = f0.squeeze()
+        pitch[-f0.shape[0]:] = f0.cpu()[:pitch.shape[0]]
+        f0 = pitch
 
-        f0 *= pow(2, f0_up_key / 12)
-        pitchf[-f0.shape[0]:] = f0.detach().cpu().numpy()[:pitchf.shape[0]]
-        f0bak = pitchf.copy()
-        f0_mel = 1127.0 * np.log(1.0 + f0bak / 700.0)
-        f0_mel = np.clip(
-            (f0_mel - f0_mel_min) * 254.0 / (f0_mel_max - f0_mel_min) + 1.0, 1.0, 255.0
-        )
-        pitch_coarse = f0_mel.astype(int)
+        if self.uv_interp:
+            uv = f0 == 0
+            if len(f0[~uv]) > 0:
+                f0[uv] = np.interp(np.where(uv)[0], np.where(~uv)[0], f0[~uv])
+            f0[f0 < self.f0_min] = self.f0_min
 
-        return pitch_coarse, pitchf
+        f0 = f0 * 2 ** (float(f0_up_key) / 12)
+
+        return f0

server/voice_changer/DiffusionSVC/pitchExtractor/PitchExtractorManager.py

@@ -1,5 +1,6 @@
 from typing import Protocol
 from const import PitchExtractorType
+from voice_changer.DiffusionSVC.pitchExtractor.CrepePitchExtractor import CrepePitchExtractor
 from voice_changer.DiffusionSVC.pitchExtractor.DioPitchExtractor import DioPitchExtractor
 from voice_changer.DiffusionSVC.pitchExtractor.HarvestPitchExtractor import HarvestPitchExtractor
 from voice_changer.DiffusionSVC.pitchExtractor.PitchExtractor import PitchExtractor
@@ -29,8 +30,8 @@ class PitchExtractorManager(Protocol):
             return HarvestPitchExtractor()
         elif pitchExtractorType == "dio":
             return DioPitchExtractor()
-        # elif pitchExtractorType == "crepe":
-        #     return CrepePitchExtractor()
+        elif pitchExtractorType == "crepe":
+            return CrepePitchExtractor()
         # elif pitchExtractorType == "crepe_tiny":
         #     return CrepeOnnxPitchExtractor(pitchExtractorType, cls.params.crepe_onnx_tiny, gpu)
         # elif pitchExtractorType == "crepe_full":

server/voice_changer/common/VolumeExtractor.py

@@ -2,6 +2,7 @@ import numpy as np
 import torch
 import torch.nn as nn
 
+
 class VolumeExtractor:
 
     def __init__(self, hop_size: float):